Source-free domain adaptation network for rolling bearing fault diagnosis

The current purpose of cross domain rolling bearing fault diagnosis is to transfer rich diagnosis features from marked source to unmarked target domain. Most domain adaptation methods require simultaneous access to both source domain data and target domain data. However, the following two shortcomings limit the development of cross-domain bearing fault diagnosis: (1) Source dataset takes up so much storage space that it is difficult to obtain. (2) Most of the source dataset contain privacy information, and there are privacy security issues. To tackle the above two shortcomings, the source-free domain adaptation network is proposed for cross-domain rolling bearing fault diagnosis. Specifically, the network mainly consists of features extractor and fault classifier. Firstly, the source domain features extractor and the source domain fault classifier are disciplinal by using source domain dataset. Then, we discard source domain dataset and keep the source domain fault classifier unchanged. Finally, the target domain features extractor is initialized by the disciplinal source domain features extractor. Using information maximization and self-supervised pseudo labels, the high-dimensional features collected by the target domain features extractor can be accurately identified by the source domain fault classifier. We use the gearbox rolling bearing dataset to verify the validity of the proposed method. The experimental result displays that this network has higher diagnosis precision compared with other traditional methods.